Abstract
Many conferences and journals use double-blind peer review to ensure the fairness of the review process. In double-blind peer review process, neither author nor reviewer identities are revealed. One interesting research question is to see whether the double-blind paper review process can indeed conceal the authors' identities. For this, we consider an authors detection problem of double-blind papers to see whether the authors of the double-blind papers can be detected with the information of their past publications. To solve the authors detection problem, we rst collect a large set of papers from arXiv. Based on the bag-of-word model, we parse these papers to extract terminologies of authors to construct a document-term matrix and an author-term matrix. By using these matrices, we propose three prediction methods to detect the authors: (i)cosine similarity, (ii)random walk on the bipartite graph, and (iii)matrix factorization, which are collaborative ltering and content-based ltering techniques. We compare the accuracy of these three methods in our experiments. Experimental results show that the cosine similarity method has highest accuracy 94%. However, the computation time of the cosine similarity method might be too long. Therefore, we use minhash to improve the efficiency of the cosine similarity method. We can see that authors who wrote more than 20 papers have more than 90% probability of being detected by our system; in other words, it is difficult to conceal authors' identities of those who published a lot of papers in the past.